Artificial intelligence is changing drug discovery almost everywhere. But in Boston, something more interesting is happening.
The city is not simply producing startups that use AI to search chemical databases faster. Greater Boston is building a much broader system in which machine learning, biology, automated experiments, protein design, clinical development, venture capital and large pharmaceutical companies increasingly work together.
That distinction matters.
AI can make a prediction in seconds. Developing a medicine still requires years of biology, experiments, manufacturing work, safety testing and clinical trials. The strongest AI drug-discovery hubs therefore need much more than great software engineers. They need laboratories, experimental scientists, hospitals, drug developers, investors, pharmaceutical partners and people who understand what can actually become a medicine.
Boston already has those pieces.
Massachusetts-headquartered biopharma companies raised $3.45 billion in venture capital during the first half of 2026, according to MassBio. That was about 25% of all U.S. biopharma venture funding during the period. The state also accounted for roughly 17% of the U.S. drug-development pipeline, while receiving 9.3% of U.S. NIH funding.
Those numbers describe the traditional strength of Massachusetts biotechnology.
The more important question for the next decade is this:
What happens when one of the world’s strongest biotechnology clusters becomes one of the places experimenting most aggressively with AI-driven science?
The answer is beginning to emerge across Cambridge, Somerville, Boston and nearby communities.
Generate Biomedicines is designing therapeutic proteins with machine learning. Nabla Bio is combining generative protein models with large-scale laboratory testing. Cellarity is using AI and multi-omics to model disease at the cell-state level. Dyno Therapeutics is using machine learning to design gene-delivery vectors. Manifold Bio is testing AI-designed proteins directly in living systems. Montai is applying multimodal AI to difficult small-molecule discovery. Relay Therapeutics has spent years integrating computational and experimental methods around protein motion. Lila Sciences is going even further by building AI systems that can participate directly in scientific experimentation.
Individually, these companies are interesting.
Together, they suggest something bigger.
Boston is beginning to look less like a collection of AI-biotech startups and more like an AI-to-medicine production system.
The Short Version: Boston Is Combining AI With the Hard Parts of Making Medicines
The easiest way to misunderstand AI drug discovery is to focus entirely on the AI.
A powerful model can suggest a protein sequence, rank a molecule or predict an interaction. But a prediction is not a medicine. Somebody still has to build the molecule, test it, understand what it does inside biological systems, improve it, manufacture it and eventually prove in human trials that it is useful.
Boston’s advantage is that many of those activities can happen inside one regional ecosystem.
Massachusetts had 113,503 biopharma jobs in 2025, even after employment fell 3.1% during a difficult industry year. The state still contains one of the world’s deepest pools of pharmaceutical researchers, computational biologists, physicians and drug-development professionals.
It also has institutions such as MIT, Harvard, the Broad Institute and major research hospitals sitting close to venture firms, startup laboratories and pharmaceutical companies.
This means an AI company in Boston can potentially recruit a machine-learning engineer, protein scientist, medicinal chemist and experienced clinical-development leader without building each professional network from scratch.
That combination is difficult to reproduce.
BostonBizScene Original Finding #1: Eight Representative Companies Have Attracted at Least $2.9 Billion in Disclosed Capital
To understand the scale of the emerging ecosystem, BostonBizScene reviewed eight Greater Boston companies in which computational or AI-driven discovery is central to the company or its scientific platform:
Generate Biomedicines, Relay Therapeutics, Lila Sciences, Cellarity, Dyno Therapeutics, Nabla Bio, Manifold Bio and Montai Therapeutics.
We then counted major publicly disclosed equity financings, IPO proceeds or explicitly committed founding capital that could be independently identified. We intentionally did not count headline milestone values from pharmaceutical partnerships because those payments may depend on future scientific or commercial events.
The result is therefore best treated as a conservative measure of visible financing rather than a complete estimate of everything invested in these businesses.
| Company | Greater Boston base | Selected publicly disclosed capital used in our analysis | Core AI/science model |
|---|---|---|---|
| Generate Biomedicines | Somerville | ~$1.09B | Generative protein therapeutics |
| Relay Therapeutics | Cambridge | ~$776M* | Computational protein dynamics and drug design |
| Lila Sciences | Cambridge | $550M | Autonomous scientific discovery |
| Cellarity | Somerville | $274M | AI, multi-omics and cell-state drug discovery |
| Dyno Therapeutics | Watertown | $109M | AI-designed gene-delivery vectors |
| Montai Therapeutics | Cambridge | $50M | Multimodal AI and small-molecule discovery |
| Manifold Bio | Boston | $40M | AI-guided protein design tested in vivo |
| Nabla Bio | Cambridge | $37M | Generative antibody and protein design |
| Combined minimum | — | ~$2.93B | — |
*For Relay, our simple comparison includes its $460 million 2020 IPO and approximately $316 million of gross proceeds reported from its May 2026 follow-on offering. It does not attempt to reconstruct every private financing or later capital event. Relay reported approximately $911 million in cash, cash equivalents and investments at the end of Q2 2026.
Generate’s figure includes its $370 million Series B, $273 million Series C, earlier Flagship backing referenced by the company and its $400 million 2026 IPO. Generate said after its 2023 Series C that it had raised nearly $700 million in equity financing since 2020; its February 2026 IPO then generated another $400 million in gross proceeds.
Lila reported $200 million of seed capital followed by a $350 million Series A, bringing total capital raised to $550 million. Cellarity reported total financing of $274 million following its Series C. Dyno publicly reported a $9 million seed and $100 million Series A. Nabla Bio has disclosed an $11 million seed and $26 million Series A. Montai began with a $50 million Flagship commitment, while Manifold Bio publicly announced a $40 million Series A.
What the $2.9 Billion Number Really Tells Us
The important insight is not that every dollar was invested because of AI.
That would be impossible to prove.
The useful conclusion is that computational drug discovery in Greater Boston is no longer confined to tiny experimental companies operating on seed capital. The region now contains AI-centered or computation-heavy companies that have progressed across several financing stages, from early venture funding through large private rounds and public markets.
The capital is also highly concentrated.
In our sample, Generate, Relay and Lila account for about 82.6% of the roughly $2.93 billion total.
Chart: Capital Concentration in the BostonBizScene Eight-Company Sample
| Company | Approx. included capital | Share of sample |
|---|---|---|
| Generate Biomedicines | $1.093B | 37.3% |
| Relay Therapeutics | $776M | 26.5% |
| Lila Sciences | $550M | 18.8% |
| Cellarity | $274M | 9.4% |
| Dyno Therapeutics | $109M | 3.7% |
| Montai Therapeutics | $50M | 1.7% |
| Manifold Bio | $40M | 1.4% |
| Nabla Bio | $37M | 1.3% |
Top three companies: 82.6% of measured capital
That concentration is strategically important.
It suggests Boston’s AI-biotech market is developing a small number of very heavily financed platform companies while still supporting younger specialists around them. Instead of every company attempting the same form of AI discovery, capital is flowing into different layers of the problem.
That is usually what a maturing technology ecosystem begins to look like.
Why Boston Has an Advantage That Pure AI Hubs Do Not Automatically Have
Boston is not necessarily the easiest place in America to build a software startup.
Labor is expensive. Laboratory facilities are expensive. Running biological experiments is expensive. Drug development can consume extraordinary amounts of capital.
Yet those apparent weaknesses contain part of Boston’s advantage.
AI drug discovery is not normal software.
When software produces a bad recommendation, engineers can often deploy another version. When a drug-development program produces a bad molecule, a company may lose months or years of experimental work.
The cost of being wrong is much higher.
That creates unusually high value for ecosystems where computational scientists sit close to experienced experimental scientists.
The Best AI Drug Companies Are Becoming Closed-Loop Systems
Look carefully at what several Boston companies say they are building.
Generate describes machine learning and experimentation as a continuous feedback loop rather than separate activities. Its platform had generated, built and tested more than 42,000 proteins, according to the company’s current website.
Nabla Bio similarly combines generative drug design with large-scale human-relevant testing. Its Cambridge team says it owns the data, AI and connected dry-lab and wet-lab systems involved in that process.
Manifold Bio goes further into living biology. It combines AI-guided protein design with high-throughput testing inside living systems rather than relying only on experiments conducted outside them.
Cellarity describes its system as a “lab-in-the-loop” approach in which experimental data continuously improves AI-driven predictions around disease cell states.
These companies are different, but the architecture is similar:
Model → design → experiment → measure → learn → redesign.
That loop may ultimately matter more than model size.
BostonBizScene Original Finding #2: The Real Moat Is Experimental Data, Not Just AI
Public discussion about AI biotechnology often focuses on algorithms.
Our review of Boston companies suggests that framing is becoming incomplete.
Among the eight companies in our sample, most describe some form of proprietary experimental, biological or clinical feedback as part of the platform. Several operate their own wet-lab infrastructure or build high-throughput experimental systems specifically designed to produce data their AI models can learn from.
That matters because public biological data has limits.
Every competitor can potentially train on widely available protein structures, scientific papers or public chemical libraries. Proprietary experimental results are different. They can reveal how molecules behave under specific conditions that are difficult or expensive for another company to recreate.
In other words, the long-term advantage may not be:
“We have the best AI model.”
It may increasingly be:
“We have built the fastest machine for producing the experimental data that makes our AI better.”
A Simple Model of the New AI-Biotech Flywheel
| Stage | Traditional approach | Emerging Boston AI-native approach |
|---|---|---|
| Define biological problem | Human research | Human research + large biological datasets |
| Propose solution | Small number of expert hypotheses | AI generates or ranks much larger design spaces |
| Build candidates | Limited experimental set | High-throughput automated creation |
| Test | Sequential lab work | Parallel, increasingly automated experiments |
| Analyze | Scientists interpret batches | Models analyze large experimental datasets |
| Learn | Knowledge stored in teams and reports | Results feed directly back into computational systems |
| Repeat | Slow project cycles | Faster design-test-learn loops |
This is where Boston’s existing biotech infrastructure becomes unusually useful.
AI needs data.
Drug-discovery AI needs good biological data.
Producing that data requires scientists, instruments, laboratory workflows, experimental design and money. Boston already had all of those before generative AI became fashionable.
AI is therefore being added to an existing discovery machine instead of trying to create that machine from zero.
The Broad Institute Helps Explain Why This Can Happen in Boston
A useful sign of where an ecosystem is going is what its scientific institutions consider worth bringing people together to discuss.
The Broad Institute’s Machine Learning in Drug Discovery Symposium spans target validation, hit identification, optimization, biologics, clinical trial design and other parts of drug discovery. Its 2026 symposium lists Isomorphic Labs President Max Jaderberg as keynote speaker.
That breadth is significant.
Machine learning is not being treated as a narrow specialty hidden in one computational department. It is increasingly relevant across multiple stages of therapeutic research.
Boston’s academic institutions also create something difficult to capture in a funding spreadsheet: constant movement between research and company formation.
Dyno traces its scientific roots to work by founder Eric Kelsic in George Church’s Harvard laboratory. Nabla Bio also emerged from work linked to the Church lab and Harvard. Manifold Bio counts Church as a co-founder.
One academic network has therefore helped seed several companies attacking different AI-biotechnology problems.
That is how clusters compound.
A professor does not need to start every company personally. Graduate students, postdocs, collaborators and experienced executives move through overlapping networks. One company’s employees can eventually become another company’s founders. Investors learn what kinds of technologies matter. Pharmaceutical companies learn where to look for partnerships.
The knowledge remains geographically sticky even when individual companies change.
Generate Biomedicines Shows What AI Drug Discovery Looks Like When It Matures
Generate Biomedicines is particularly important because it has moved far beyond the idea-stage version of generative biology.
The Somerville company uses machine learning to design protein therapeutics. It describes its system as a continuous loop in which machine learning powers protein design and experiments continually improve the platform.
The company raised $370 million in 2021 and another $273 million in 2023 before completing a $400 million IPO in February 2026.
More importantly, the platform is being tested through actual drug development.
Generate announced in December 2025 that it planned global Phase 3 studies for GB-0895, an AI-engineered long-acting antibody for severe asthma. That is qualitatively different from showing that an algorithm can predict a protein structure inside a research benchmark.
Clinical development is where the strongest claims about AI drug discovery eventually have to face biology.
Why That Matters for Boston
For years, one of the biggest questions around AI drug discovery has been whether faster computational design translates into clinically useful drugs.
Nobody should assume that one successful program proves the entire AI-drug-discovery thesis.
Drug development remains difficult, and clinical trials can fail for reasons that no discovery model can fully eliminate.
But companies progressing computationally designed programs toward late-stage development change the discussion. The argument is no longer purely theoretical.
Boston now has companies attempting to connect AI-generated molecular design to the same development system used by the global pharmaceutical industry.
That is an important milestone for the region.
Relay Therapeutics Shows That Computational Drug Discovery Started Before the Generative-AI Boom
It is tempting to describe AI drug discovery as something that began when generative AI exploded into public view after 2022.
Boston’s history is older.
Relay Therapeutics built its scientific approach around understanding the movement and behavior of proteins and combining computational methods with experimental drug discovery.
The Cambridge company completed a $460 million IPO in 2020. In May 2026 it completed another large public financing; by the end of the second quarter, Relay reported approximately $911 million in cash, cash equivalents and investments.
Relay also has clinical programs rather than merely computational prototypes.
This distinction helps explain an important feature of Boston’s ecosystem.
The region does not need to choose between “old biotech” and “new AI.”
The more interesting companies combine them.
Medicinal chemistry still matters. Structural biology still matters. Assay design still matters. Clinical development still matters.
AI becomes another powerful layer inside that system.
Nabla Bio Shows Why Protein Design Is Becoming a Boston Specialty
Nabla Bio provides a cleaner example of the current generative-AI generation.
The Cambridge company combines generative protein design with high-throughput wet-lab experimentation and is focused heavily on antibody therapeutics.
Its scientific lineage also reflects Boston’s academic flywheel. Nabla was founded by researchers connected to George Church’s Harvard lab, and investor Radical Ventures describes the founders as early pioneers of protein language models.
Nabla raised an $11 million seed round in 2021 and a $26 million Series A in 2024, bringing disclosed equity financing from those two rounds to $37 million.
But its commercial partnerships may be even more revealing.
When Nabla announced the Series A, it also disclosed collaborations with AstraZeneca, Bristol Myers Squibb and Takeda whose potential upfront and milestone value exceeded $550 million. In 2025 it announced another Takeda collaboration involving double-digit millions in upfront and research payments and potential success-based payments exceeding $1 billion.
These headline values should not be confused with money already received. Milestone-based pharmaceutical agreements often pay substantial amounts only if programs achieve defined future goals.
Still, the existence of the partnerships tells us something important.
Large pharmaceutical companies are not merely watching Boston’s AI-protein companies.
They are using them.
Manifold Bio Reveals the Next Bottleneck: Testing What AI Designs
Imagine an AI model becomes dramatically better at designing proteins.
At first that sounds like the problem has been solved.
Then another problem appears.
If a system can generate one million plausible designs, how do you experimentally test them?
That is where companies such as Manifold Bio become strategically interesting.
The Boston company has developed technologies for testing large numbers of biologic designs and measuring their behavior in living systems. In March 2026, Manifold and NVIDIA reported an experiment testing one million protein-binder designs against 127 targets, producing more than 100 million protein-protein interaction measurements.
This changes the AI drug-discovery problem.
When model generation becomes cheap, measurement becomes scarce.
BostonBizScene Original Finding #3: AI Is Moving the Bottleneck From Generation to Validation
Traditional discovery often struggles because generating plausible drug candidates is difficult.
Generative AI can dramatically increase the number of candidates scientists can propose.
But biology does not care how quickly those candidates were generated.
They still need to be tested.
This means the economic bottleneck can move downstream.
| AI capability improves | New bottleneck created |
|---|---|
| Faster molecule generation | More compounds need validation |
| Better protein design | More functional testing is required |
| Larger virtual libraries | Experimental throughput becomes limiting |
| Better biological prediction | Human-relevant validation becomes more valuable |
| Autonomous hypothesis generation | Laboratory automation becomes more important |
Boston may benefit precisely because it is good at the expensive side of that table.
The region’s advantage is not simply the ability to generate candidates.
It is the ability to build experimental infrastructure around those candidates.
Manifold’s Roche partnership supports that interpretation. In November 2025, Roche agreed to pay Manifold $55 million upfront as part of a collaboration around next-generation brain shuttles, with potential future research, development and commercial milestones exceeding $2 billion.
Again, milestone value is not guaranteed revenue.
What matters is that a major pharmaceutical company is paying meaningful upfront capital to access the platform.
Dyno Therapeutics Shows That AI Drug Discovery Is Bigger Than Designing Drugs
Drug discovery conversations often focus on designing the therapeutic molecule.
But sometimes the harder problem is getting the medicine to the right place.
Dyno Therapeutics is applying machine learning to the design of AAV capsids, the protein shells used in certain gene-therapy delivery systems.
The company was founded in 2018, raised a $9 million seed round and later completed a $100 million Series A led by Andreessen Horowitz.
Its history also shows how a local AI-biotech company can plug into the global pharmaceutical system.
Dyno has worked with Novartis, Sarepta, Roche and Astellas. In 2024, Roche expanded its relationship with Dyno through an agreement carrying a $50 million upfront payment and potential milestones exceeding $1 billion.
The broader lesson is important.
AI-powered drug discovery is not one market.
It includes molecule design, protein design, target identification, delivery technology, disease modeling, experiment planning and eventually clinical-development tools.
Boston has companies working across several of these layers.
That diversity makes the cluster more resilient.
Cellarity Shows Another Direction: Use AI to Understand the Cell, Not Just the Molecule
Most traditional drug discovery begins with a target.
Researchers identify a protein believed to play an important role in disease and search for a molecule that changes its activity.
Cellarity approaches the problem differently.
The Somerville company uses high-dimensional biological data and AI to understand disease at the level of cell states. Its platform combines multi-omics, single-cell information and predictive modeling to identify interventions capable of moving dysfunctional cells toward healthier states.
That matters because AI may change not only how drugs are designed, but also what scientists choose to design drugs against.
Cellarity raised $123 million in a Series B and another $121 million in its Series C, bringing total financing reported by the company to $274 million by 2022.
Its work with Novo Nordisk also illustrates the relationship between platform companies and established drug makers. An expanded MASH collaboration announced in 2024 could generate up to $532 million across research, development and commercial payments if specified conditions are achieved, in addition to royalties.
The pattern is becoming familiar.
Boston startup builds unusual discovery platform.
Large pharmaceutical company supplies disease knowledge, development scale or commercial reach.
The two meet somewhere between software and therapeutics.
Montai Shows Why AI Is Expanding the Search Space
Another important change is happening on the chemistry side.
Traditional pharmaceutical companies can search enormous chemical libraries, but the theoretically possible universe of molecules is far larger than anything scientists can physically screen.
AI gives researchers a way to navigate more of that space.
Cambridge-based Montai Therapeutics says its CONECTA platform uses multimodal foundation models trained on more than one billion curated and proprietary data points. Its wider platform works across a library of more than three billion compounds derived from diverse chemistry and is focused on difficult inflammatory disease biology.
Flagship committed an initial $50 million when it unveiled the company in 2022.
Montai reports that its platform has produced drug candidates roughly 24 months from hit identification to candidate selection, although that is a company-reported platform metric and should not be interpreted as an industry-wide benchmark.
Still, the direction is clear.
AI allows drug companies to ask a different question.
Instead of asking, “Which molecule in our existing library works?”
They can increasingly ask, “What part of chemical or biological design space have we barely explored?”
That is a much larger opportunity.
Lila Sciences Could Push Boston From AI Drug Discovery Toward AI-Driven Science
Lila Sciences is the most expansive example in this group because the company’s ambition goes beyond drug discovery.
The Cambridge company is building AI systems designed to participate directly in scientific research. It has paired AI with automated laboratories and describes its goal as building a general platform for autonomous science across life sciences, chemistry and materials.
It has also attracted unusual financial backing.
Lila announced $200 million in seed capital and later closed a $350 million Series A, bringing total capital raised to $550 million.
The scientific model is especially important.
A conventional AI system learns mainly from existing information.
An autonomous scientific system could potentially decide what experiment would provide the most useful new information, execute that experiment through laboratory automation, examine the result and choose what to test next.
That creates another loop:
Reason → experiment → observe → update → reason again.
Lila says its platform has already tested more than 950,000 mRNA sequences in living cells, using the resulting measurements to improve its scientific reasoning models. The company also reported work applying the platform to in-vivo CAR-T design.
Those are company-reported results rather than independent proof that general autonomous science has been solved.
But they show why the Boston ecosystem may soon stretch beyond the phrase “AI drug discovery.”
The bigger market could become AI-operated research itself.
BostonBizScene Original Finding #4: Half of Our Representative Companies Are Based in Cambridge
The eight-company sample also shows how geographically concentrated this emerging sector remains.
| Location | Companies in sample | Share |
|---|---|---|
| Cambridge | 4 | 50% |
| Somerville | 2 | 25% |
| Boston | 1 | 12.5% |
| Watertown | 1 | 12.5% |
The Cambridge companies in our sample are Relay Therapeutics, Nabla Bio, Montai Therapeutics and Lila Sciences.
Generate and Cellarity are headquartered in Somerville. Manifold Bio is in Boston, while Dyno moved its headquarters to Watertown.
This should not be treated as a complete census of Boston AI-biotech companies.
It is a focused sample.
But even that small dataset makes an important point: Boston’s AI-drug-discovery ecosystem is really a Greater Boston corridor, not a downtown Boston phenomenon.
That corridor follows existing biology.
Why Kendall Square Still Matters in the AI Era
For normal software, geographic proximity has become less important in many cases.
A team can build cloud software across several countries.
Biotechnology behaves differently.
Experiments happen somewhere.
Robotic laboratories happen somewhere.
Cell cultures happen somewhere.
Scientists still need access to instruments, samples and controlled laboratory environments.
That makes physical research clusters unusually durable.
A computational biologist can work remotely for part of the week. A high-throughput laboratory cannot move into a Zoom meeting.
This is one reason the Boston-Cambridge cluster could remain valuable even as AI makes knowledge work more distributed.
The physical infrastructure still matters.
Boston’s Temporary Lab-Space Weakness Could Become an AI-Biotech Advantage
The Massachusetts biotechnology downturn produced an uncomfortable real-estate problem.
MassBio’s 2025 snapshot reported that the state had reached approximately 63.2 million square feet of laboratory and biomanufacturing inventory, while laboratory vacancy rates reached 22.9% in Cambridge and 38.3% in Boston.
For landlords, high vacancy is painful.
For new laboratory-heavy startups, it can create opportunity.
During a boom, an AI-biology company may have to fight established pharmaceutical businesses for expensive laboratory space. During a softer market, founders may have more choices, landlords have more incentive to negotiate and companies can potentially expand without waiting years for new buildings.
That does not make Boston cheap.
It does make the physical bottleneck somewhat less severe than it was at the top of the biotech real-estate cycle.
For an AI-native biotech trying to build a large experiment-generation system, that matters.
The Capital Market Is Sending a Mixed but Useful Signal
Boston’s biotech environment is not universally booming.
Employment contracted in 2025.
Seed-stage funding became more difficult.
MassBio reported that the average seed financing declined to about $4.65 million in the first half of 2026, down from $7.65 million, even while the average Series A increased to $79.6 million.
That creates a more selective market.
Capital is available, but investors appear increasingly willing to concentrate large amounts behind companies they believe have strong evidence, platforms or clinical potential.
For AI-biotech founders, the lesson is important.
Having “AI” in a pitch deck is not enough.
The companies raising serious capital increasingly have something harder to reproduce: proprietary biology, laboratory systems, meaningful experimental data, drug-development programs or pharmaceutical partnerships.
BostonBizScene Original Finding #5: The Ecosystem Is Balanced Between Designing Molecules and Redesigning Discovery
We also classified the eight-company sample by the main scientific problem each company is trying to solve.
The categories are simplified because several companies span more than one area.
| Primary area | Representative companies | Count |
|---|---|---|
| Protein/biologic design | Generate, Nabla Bio, Manifold Bio | 3 |
| Small-molecule or disease-state discovery | Relay, Cellarity, Montai | 3 |
| Gene-delivery design | Dyno Therapeutics | 1 |
| Autonomous science infrastructure | Lila Sciences | 1 |
The interesting result is not the exact 3-3-1-1 split.
It is the absence of a single dominant technical model.
Boston is not producing eight startups that all promise to generate small molecules faster.
Some are designing proteins. Some are changing how disease is represented computationally. One is redesigning delivery vectors. Another is building an autonomous scientific system.
That diversity may become one of the region’s biggest defenses against commoditization.
If foundation models make basic molecular generation easier for everyone, Boston companies can compete somewhere deeper in the stack.
Pharma Partnerships Are Becoming a Form of Validation
Venture capital can tell us what investors believe.
Pharmaceutical partnerships tell us something different.
They show what companies responsible for actual drug pipelines are willing to test, license or fund.
Several Boston companies in this article have already formed relationships with major pharmaceutical businesses.
| Boston-area company | Selected external partner examples |
|---|---|
| Nabla Bio | AstraZeneca, Bristol Myers Squibb, Takeda |
| Dyno Therapeutics | Roche, Novartis, Sarepta, Astellas |
| Cellarity | Novo Nordisk |
| Manifold Bio | Roche |
| Generate Biomedicines | Amgen; MD Anderson collaborations |
| Quotient Therapeutics* | Pfizer, GSK-related programs |
*Quotient is outside our eight-company financing sample but provides another example of the wider Cambridge ecosystem. Flagship initially committed $50 million to the company, which uses somatic genomics to uncover links between genes and disease.
These partnerships matter because large drug companies already possess tremendous internal scientific capacity.
They do not need startups simply to run ordinary experiments.
The startup generally needs to offer something difficult to reproduce internally: unusual data, faster experimentation, differentiated models, novel biology or access to a new technical platform.
That is why pharmaceutical adoption may be a more useful signal than AI publicity.
The Big Shift: AI Is Becoming Part of the Scientific Instrument
The first generation of biotechnology software mostly helped people manage information.
The next generation helped scientists analyze information.
The emerging generation increasingly helps decide what to create and test.
That is a much bigger shift.
A laboratory traditionally contains physical instruments: sequencers, microscopes, mass spectrometers, liquid-handling robots and many others.
AI is beginning to function like another scientific instrument.
It can compress enormous datasets into usable representations. It can suggest molecules a scientist might never manually design. It can identify patterns across experiments. In more advanced systems, it can potentially choose the next experiment.
The consequences extend far beyond productivity.
When researchers can explore more ideas per dollar, the kinds of questions they can afford to ask change.
Why Boston Could Win Even If the Best Foundation Models Come From Somewhere Else
Boston does not need to build every foundational AI model to remain important in AI-driven medicine.
That may be one of the most important strategic points for the region.
General-purpose AI infrastructure may increasingly come from companies such as NVIDIA, OpenAI, Anthropic, Google DeepMind and other large computing organizations.
Boston can specialize higher up the value chain.
Its companies can combine foundation models with proprietary experiments, disease knowledge, biological measurements and therapeutic-development systems.
Manifold Bio’s collaboration with NVIDIA illustrates the model. NVIDIA provided a generative protein-design model, while Manifold supplied a system capable of experimentally validating designs at enormous scale.
That relationship suggests a possible future architecture for the industry.
Big technology companies provide increasingly powerful general models.
Boston biotech companies turn those models into biological discovery engines.
Pharmaceutical companies turn promising discoveries into globally distributed medicines.
Each layer creates value.
What Founders Should Learn From Boston’s Strongest AI-Biotech Companies
Founders entering this market should resist the temptation to start with the words “AI platform.”
Start with the bottleneck.
What experiment takes too long?
What biological question produces weak data?
What class of protein remains hard to design?
What disease mechanism is poorly represented by existing targets?
What part of chemistry remains difficult to search?
What discovery step cannot scale because scientists still perform it sequentially?
The stronger Boston companies tend to connect AI to a specific scientific constraint.
Build the Data Loop Before Building the Marketing Story
A model without proprietary feedback can become replaceable.
A company with a continually growing dataset produced by its own experiments becomes harder to copy.
The ideal system improves through use.
Every experimental cycle should create information that helps the next cycle.
That is a much stronger moat than claiming access to the newest language model.
Design Around Validation From Day One
The value of an AI prediction depends heavily on the cost and quality of validating it.
Founders should therefore design computational and experimental systems together.
If a model can generate 100,000 candidate molecules but the company can test only ten each quarter, most of that computational power is wasted.
The measurement system becomes part of the product.
Decide Whether You Are Building a Drug Company or a Discovery Company
This distinction becomes strategically important as companies mature.
A platform company can partner broadly and generate revenue from pharmaceutical relationships.
A therapeutics company can capture much more value if its own drugs succeed, but clinical development requires far greater capital and risk.
Some Boston companies are attempting to do both.
That strategy can work, but management teams need to understand the trade-off.
Building a world-class AI platform and a world-class clinical organization are two expensive jobs.
What Investors Should Look For
The easiest AI-biotech diligence question is also one of the least useful:
“How advanced is your AI?”
A better discussion starts with evidence.
How many designs has the system generated?
How many were actually built?
How many were tested?
What percentage showed useful activity?
How does performance compare with a credible baseline?
How quickly does experimental feedback return to the model?
Does the platform improve as proprietary data grows?
Has an external pharmaceutical team reproduced or paid for the value?
These questions move evaluation from storytelling to measurable scientific productivity.
A Practical AI Drug-Discovery KPI Dashboard
| KPI | Why it matters |
|---|---|
| Designs generated per cycle | Shows computational throughput |
| Designs experimentally tested | Shows real validation capacity |
| Hit rate versus baseline | Tests whether AI improves outcomes |
| Time from design to experiment | Measures loop speed |
| Cost per useful experimental result | Connects science to economics |
| Proprietary measurements generated | Measures growth of data advantage |
| Repeatability across programs | Distinguishes platform from one success |
| Partner renewals or expansions | Provides outside validation |
| Programs entering IND-enabling work | Measures movement toward development |
| Clinical programs | Shows whether discovery translates downstream |
No single KPI proves platform quality.
Together, however, they can tell investors much more than a model demonstration.
What Pharmaceutical Companies Should Do Differently
Large pharmaceutical companies should avoid treating AI partnerships as innovation theater.
The useful question is not whether a company has an AI collaboration.
The question is whether the collaboration changes the economics or capability of discovery.
Can a partner reach previously inaccessible targets?
Can it produce higher-quality starting molecules?
Can it test biology at larger scale?
Can it shorten experimental cycles?
Can it create proprietary data the pharmaceutical company does not possess?
If the answer is yes, the relationship may create strategic value even before a drug reaches the market.
Boston provides a unusually dense market in which pharmaceutical companies can test different versions of that thesis.
What Boston Policymakers and Ecosystem Builders Should Protect
Boston’s position is strong, but it is not guaranteed.
The MassBio employment data is a useful warning.
Massachusetts biopharma employment fell 3.1% in 2025, including a 3.9% decrease in R&D jobs.
A region can lead scientifically while gradually becoming too expensive or difficult for younger companies to build in.
AI may intensify that tension.
The most advanced AI-biotech companies need expensive GPUs, expensive scientists and expensive laboratory systems at the same time.
That means Boston should think about competitiveness across the complete cost stack.
Lab availability matters.
Early-stage funding matters.
Connections between universities and startups matter.
Workforce development matters.
Compute access increasingly matters.
Policies that help researchers translate academic work into companies matter.
The ecosystem should not focus only on attracting mature companies with hundreds of employees.
The next Generate or Dyno may begin with five people and an experiment that looks strange.
Boston’s Biggest Competitive Threat Is Not San Francisco Alone
It is easy to frame this as Boston versus San Francisco.
That comparison is incomplete.
AI-powered drug discovery is becoming global.
The United Kingdom has major AI-biology activity around organizations such as Google DeepMind and Isomorphic Labs. California has deep machine-learning talent, biotechnology companies and venture capital. China is rapidly expanding both biotechnology research and its drug-development pipeline.
MassBio’s 2026 report explicitly highlighted China’s growing position in biotechnology as an important competitive issue for Massachusetts.
Boston therefore competes on several fronts at once.
Its advantage is not that nobody else can build good AI models.
Its advantage is ecosystem density.
The challenge is keeping that density productive enough to justify Boston’s costs.
The Most Important Race May Be Experimental Throughput
A decade ago, AI drug discovery was often framed as a prediction problem.
Who could predict the best molecule?
That may no longer be the central question.
If many companies gain access to excellent foundation models, the competitive advantage shifts toward the ability to generate high-quality proprietary evidence.
Who can run more useful experiments?
Who can make those experiments cheaper?
Who can automate them?
Who can measure biology at greater scale?
Who can use experimental failures as training data?
Who can close the model-to-lab loop fastest?
This is where Boston’s unusual mix of biotechnology and AI becomes powerful.
BostonBizScene Original Finding #6: The Winning Architecture Is Becoming AI + Wet Lab + Pharma
Across the companies examined for this article, three capabilities repeatedly appear.
Layer One: Computational Intelligence
The company needs models capable of navigating biological or chemical complexity.
That might mean protein-language models, generative models, cell-state models, multimodal systems, molecular simulations or scientific agents.
Layer Two: Experimental Intelligence
The company needs a way to determine whether the computational output is biologically useful.
Generate builds and tests proteins.
Nabla links design to experiments.
Manifold measures designs in living systems.
Dyno combines machine learning with high-throughput biological measurements.
Cellarity uses experimental data inside its discovery loop.
Lila is attempting to automate more of the scientific cycle itself.
Layer Three: Development Intelligence
Eventually somebody must know how to turn a scientifically interesting result into a medicine.
This includes toxicology, manufacturing, regulatory work, clinical development and commercial strategy.
Boston’s established pharmaceutical and biotech community provides unusually deep access to people who have performed these jobs before.
Put those layers together and a useful model emerges:
AI creates search power.
Experiments create truth.
Drug-development experience creates medicines.
Boston has all three.
The Next Boston AI-Biotech Wave Could Look Very Different
The current leaders will not necessarily define the entire market.
Several new categories are likely to grow.
Autonomous Labs
Instead of waiting for scientists to manually move every project between computational and laboratory teams, software agents could schedule experiments, interpret measurements and choose follow-up experiments.
Lila represents one ambitious version of this future.
AI-Native Clinical Development
Discovery receives most of the attention, but trials consume enormous time and capital.
AI could increasingly influence patient selection, trial-site decisions, biomarkers, protocol design and clinical operations.
Biology Foundation Models
Models trained across DNA, RNA, proteins, cells, imaging and clinical information may create more unified representations of disease.
That could move drug discovery away from isolated targets toward more complete biological systems.
AI-Designed Delivery Systems
Dyno demonstrates that the design problem does not stop with the therapeutic payload.
Gene therapy, RNA medicine and other advanced modalities need better delivery.
AI could become central to engineering those delivery systems.
AI-Driven Experimental Infrastructure
If candidate generation keeps accelerating, companies that help laboratories test those candidates faster could become extremely valuable.
Manifold points toward this opportunity.
The picks-and-shovels layer of AI biology may eventually become as strategically important as the model layer.
The Economics of Drug Discovery Could Change in a Subtle Way
AI does not need to cut the complete drug-development process from twelve years to two years to create enormous value.
That expectation is probably too simplistic anyway.
Many parts of clinical development are constrained by biology, patient recruitment and regulatory requirements rather than computation.
A more realistic economic effect is cumulative.
Suppose AI helps scientists reject weak ideas earlier.
Suppose better protein design raises the quality of starting candidates.
Suppose automated experiments allow teams to evaluate ten times more variants.
Suppose predictive models reduce unnecessary synthesis.
Suppose disease models help researchers choose stronger biological hypotheses.
Each improvement may affect only one stage.
Together they can change the economics of the entire portfolio.
The biggest financial benefit may therefore come not from making every successful drug dramatically faster, but from spending less money pursuing ideas that were never going to work.
Why Failure Data Could Become Surprisingly Valuable
Drug discovery creates huge amounts of failure.
Most compounds do not become medicines.
Traditionally, failure is treated mainly as cost.
AI can partly change that.
If an experimental system records failures consistently, unsuccessful designs become training data.
A protein that does not bind teaches the system something.
A molecule with poor properties teaches the system something.
A delivery vector that fails to reach the desired tissue teaches the system something.
This is why integrated AI-biotech platforms can compound.
The laboratory is not only validating the model.
It is manufacturing future training data.
A well-designed company therefore gets smarter even when an individual experiment fails.
That could become one of the deepest differences between traditional project-based drug discovery and AI-native discovery systems.
Why Boston’s Biotech Downturn Does Not Automatically Break the Thesis
Massachusetts biotechnology has gone through a difficult adjustment.
Companies have reduced headcount.
Lab vacancies have increased.
Early-stage capital has become more selective.
Those facts should not be hidden behind an optimistic AI story.
But they do not necessarily contradict the AI-hub thesis.
In fact, some parts of the correction can accelerate restructuring.
Experienced scientists become available.
Laboratory space becomes easier to obtain.
Investors demand clearer proof.
Companies can no longer raise giant rounds simply by attaching fashionable language to ordinary science.
Strong AI-biotech companies may emerge from that environment more disciplined.
The important metric will not be how many startups use the letters “AI.”
It will be how many produce better science.
A Strategic Scorecard for Boston’s AI Drug-Discovery Position
Rather than ranking Boston against other cities using arbitrary scores, it is more useful to identify the observable conditions supporting the cluster.
| Ecosystem factor | Evidence visible in Greater Boston |
|---|---|
| Biopharma capital | $3.45B raised by MA-headquartered biopharma companies in H1 2026 |
| National funding share | Approximately 25% of U.S. biopharma VC in H1 2026 |
| Drug pipeline | Massachusetts companies account for roughly 17% of U.S. pipeline |
| Biomedical research | Massachusetts received 9.3% of NIH funding |
| Experienced workforce | 113,503 biopharma jobs in 2025 despite contraction |
| Academic research | Harvard, MIT, Broad and major hospital research networks |
| AI-native drug companies | Multiple companies across proteins, small molecules, delivery and autonomous science |
| Pharma validation | Partnerships involving Roche, Takeda, AstraZeneca, BMS, Novo Nordisk and others |
| Experimental infrastructure | Large existing lab footprint plus AI-native wet-lab systems |
| Public-market validation | Generate and Relay have accessed public equity markets |
MassBio data supports the statewide capital, employment, pipeline and NIH figures.
The combination is more important than any individual line.
Plenty of places have computer scientists.
Plenty have universities.
Plenty have biotechnology companies.
Plenty have venture capital.
Very few have all of them concentrated closely enough for talent, companies and ideas to move between them continuously.
What Could Stop Boston From Becoming an Even Bigger AI-Drug Hub?
Three risks deserve particular attention.
Cost Could Push the Earliest Companies Elsewhere
Mature biotech companies can tolerate Boston’s cost structure better than two founders working on an unusual scientific idea.
If seed rounds continue shrinking while laboratory and talent costs remain high, younger teams may start elsewhere.
That would weaken the bottom of the ecosystem’s company-formation funnel.
AI Talent Has More Choices Than Biotech Talent Used To
A machine-learning researcher interested in biology can work for a biotech company.
But that same person may also choose a frontier AI laboratory, major technology company or quantitative finance firm.
Biotechnology must therefore compete for technical talent on compensation, computing resources, research freedom and impact.
Boston Cannot Assume Historic Biotech Leadership Guarantees AI Leadership
Existing strength creates an advantage.
It does not create ownership of the future.
The best AI-biology companies could emerge anywhere that combines computing, capital and science effectively.
Boston needs to keep building rather than merely defending its history.
What Boston Businesses Outside Biotech Should Pay Attention To
AI-powered drug discovery may sound like a narrow topic for scientists.
Its local economic effects could become much broader.
These companies need specialized real estate.
They need laboratory automation.
They need cloud and compute infrastructure.
They need recruiters.
They need legal services.
They need intellectual-property work.
They need clinical consultants.
They need manufacturing partners.
They need specialized data infrastructure and cybersecurity.
They need construction firms that understand advanced laboratories.
A successful AI-biotech cluster therefore produces opportunities around the core companies.
Boston service businesses should pay attention not just to who raises funding, but where spending moves after the round closes.
An AI-biotech company building a larger experimental engine may generate business for dozens of supporting firms.
The Metric Boston Should Watch Next
Funding is easy to measure.
Company creation is easy to count.
The more useful long-term metric may be translation.
How many AI-designed or AI-discovered programs reach clinical testing?
How many pharmaceutical partnerships expand after their first projects?
How many platforms repeatedly produce drug candidates rather than a single success?
How many discovery cycles become materially faster or cheaper?
How many companies generate proprietary experimental datasets at a scale that competitors cannot easily match?
Those questions will tell us whether Boston is building durable scientific capability or simply experiencing another technology investment cycle.
The Bigger Story: Boston Is Turning Biology Into an Engineering Loop
The most important development is not that scientists are using AI.
Scientists have used computational tools for decades.
The bigger change is that parts of biology are becoming iterative in a way that looks increasingly like engineering.
Design something.
Build it.
Measure it.
Feed the result back into the system.
Improve the design.
Repeat.
Software engineers have worked this way for years because software is easy to copy, test and modify.
Biology is harder.
Cells are messy.
Animals are complicated.
Human disease is far more complicated than either.
But when automation lowers the cost of experiments and AI increases the speed of design, the loop gets tighter.
Boston is unusually well placed to benefit because the city already built much of the physical and human infrastructure required for the experimental half of that loop.
Final Takeaway: Boston’s Advantage Is Not AI Alone
Boston is becoming a global center for AI-powered drug discovery because it is combining two ecosystems that are extremely difficult to build separately.
One is artificial intelligence.
The other is biotechnology.
The strongest local companies do not treat AI as magic software sitting above biology. They connect algorithms to experiments, proprietary data, drug-development expertise and pharmaceutical partners.
Our BostonBizScene analysis of eight representative companies found at least $2.93 billion in selected publicly disclosed or committed capital, with Generate Biomedicines, Relay Therapeutics and Lila Sciences accounting for roughly 82.6% of the total.
More important than the money is what those companies are building.
Generate is connecting generative protein design to clinical development.
Nabla is combining protein models with high-throughput experiments.
Manifold is attacking the experimental validation bottleneck.
Dyno is applying AI to gene-delivery design.
Cellarity is modeling disease at the cell-state level.
Montai is exploring difficult chemical space.
Relay demonstrates the maturity of computation-driven therapeutic development.
Lila is attempting to automate larger parts of the scientific method itself.
The pattern is hard to ignore.
Boston’s AI-biotech future will probably not be defined by one giant model sitting in a Cambridge office.
It will be defined by thousands or millions of cycles connecting models with real biology.
That is the strategic advantage Boston already understands.
The city does not just know how to build software.
It knows how to turn scientific ideas into experiments, experiments into companies and, occasionally, companies into medicines.
AI could make that machine move much faster.
